You are on page 1of 24

Board of Regents of the University of Wisconsin System

Marriage, Motherhood, and Wages


Author(s): Sanders Korenman and David Neumark
Source: The Journal of Human Resources, Vol. 27, No. 2 (Spring, 1992), pp. 233-255
Published by: University of Wisconsin Press
Stable URL: http://www.jstor.org/stable/145734
Accessed: 13-11-2015 07:17 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact support@jstor.org.

Board of Regents of the University of Wisconsin System and University of Wisconsin Press are collaborating with
JSTOR to digitize, preserve and extend access to The Journal of Human Resources.

http://www.jstor.org

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
Marriage, Motherhood, and Wages

Sanders Korenman
David Neumark

ABSTRACT
We explore several problems in drawing causal inferences from
cross-sectional relationships between marriage, motherhood,
and wages. We find that heterogeneity leads to biased estimates
of the "direct" effects of marriage and motherhood on wages
(i.e., effects net of experience and tenure); first-difference esti-
mates reveal no direct effect of marriage or motherhood on
women's wages. We also find statistical evidence that experi-
ence and tenure may be endogenous variables in wage equa-
tions; instrumental variables estimates suggest that both
ordinary least squares cross-sectional and first-difference esti-
mates understate the direct (negative) effect of children on
wages.

I. Introduction
Cross-sectional studies find little association between a
woman's marital status and her wage rate, but often a negative relation-
Sanders Korenmanis an assistant professor of economics and public affairs at Princeton
Universityand is a faculty associate of the Officeof PopulationResearch at Princeton
Universityand a research affiliate of the National Bureauof EconomicResearch
(NBER). David Neumarkis an assistant professor of economics at the Universityof
Pennsylvaniaand a research affiliateof the NBER. The authorsthankMcKinleyBlack-
burn, David Bloom, RichardFreeman, ClaudiaGoldin,LawrenceKatz, Bruce Meyer,
anonymousreferees, and seminarparticipantsat the National Bureauof EconomicRe-
search Labor Studies Program, Princeton University,and the Universityof South Caro-
linafor helpful comments, and Elaina Rose for researchassistance. The data used in
this article can be obtainedfrom David Neumarkat the following address:Department
of Economics, 518 McNeil, Universityof Pennsylvania,3718 Locust Walk,Philadelphia,
PA 19104.
[Submitted May 1989; accepted March 1991]
THE JOURNAL OF HUMAN RESOURCES * XXVII * 2

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
234 The Journal of Human Resources

ship between childrenand wages. The negativerelationshipbetween chil-


dren and wages is reduced and sometimes eliminated by inclusion in
wage equations of detailed controls for "laborforce attachment"such as
experience and tenure.
There are, however, a number of reasons to be cautious in drawing
causal inferences from these cross-sectional relationships. First, labor
market experience and tenure may be endogenous if labor supply is re-
sponsive to wages. Because the estimated effect of childrenon wages is
sensitive to the inclusion of controls for experience and tenure, it is im-
portant to explore whether experience and tenure are in fact exogenous
variables in wage equations. Second, economic theories of fertility and
marriage(e.g., Becker 1981, Butz and Ward 1979, Easterlin 1980), sug-
gest that maritalstatus and numberof childrenmay also be endogenous
with respect to wages. Third, estimatedwage effects of marriageor chil-
dren may be biased by unmeasuredheterogeneity:women may be se-
lected or may self-select into different maritalor fertility states on the
basis of unmeasuredcharacteristicsthat are correlatedwith wages (e.g.,
"careerorientation"),even if marriageor fertilityare not directlyrespon-
sive to wages. Finally, bias could result if, among marriedwomen or
women with children, those with high wages tend to select into employ-
ment (i.e., the standardproblemof sample selection bias).
Previous researchershave recognizedsome of these potentialproblems
in interpretingcross-sectional relationships between wages, marriage,
and children,but have not attemptedto evaluatethe empiricalimportance
of each of them in a single data set. This paper presents evidence on the
magnitudesof these biases, assesses the sensitivity of the estimated ef-
fects to alternativeapproachesto eliminatingbias, and attemptsto arrive
at unbiased estimates of the effects of marriageand childrenon wages.
The effects of marriageand motherhood on wages are of particular
interest due to their relationto male-femalewage differentials.For exam-
ple, Becker (1985) has hypothesized that a portion of male-femalewage
differentials is attributable to gender-role specialization by married
women and men. In particular,he has arguedthat the "hourly earnings
of single women [should] exceed those of marriedwomen even when
both work the same numberof hours and have the same marketcapital
because child care and other household responsibilitiesinduce married
women to seek more convenient and less energyintensivejobs" (p. S54).1
Therefore,the empiricalanalysis of wage differentialsbetween single and
marriedwomen, and women with and without children,can shed light on
the wage effects of gender-rolespecialization.

1. Fuchs (1989) makes a similar argument.

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
Korenman and Neumark 235

II. Empirical Studies of Marriage, Children, and


Women's Wages

Using the 1976Panel Study of Income Dynamics, in multi-


ple regression analyses with a standard set of human capital controls,
Hill (1979) finds that married,white women with a spouse present earn
more than their never-marriedcounterparts,but less than divorced, sepa-
rated, or widowed women. However, these differences are small and
not statistically significant.Controllingfor maritalstatus, the numberof
childrenpresent is significantlynegatively related to earnings. But when
Hill adds detailed measures of labor force attachmentand interruptions,
the negative association between childrenand wages becomes small and
insignificant(less than 1 percent per child).2'3
Focusing exclusively on the association between children and wages,
Moore and Wilson (1982)examine a cross-section of marriedwomen aged
35-49 in 1972who were full-timeworkers, from the NLS MatureWomen
file. Controllingfor a wide variety of worker characteristics, they find
that marriedwomen with three or more children earn about 11 percent
less per hour than marriedwomen without children, but that there are
small and statistically insignificantdifferences among women with zero,
one, or two children.
Goldin and Polachek (1987)examine the 1/1,000 Public Use Sample of
the 1980 U.S. Census of Populationand find substantialand significant
annual earnings differentialsfavoring never-marriedwomen. However,
substituting a variable called the "Expected Human Capital Stock"
(which varies across individualswith differences in expected lifetime la-
bor force participation)for education and experience variables reduces
the marriagedifferentialto about 38 percent of its original size.4 Goldin
and Polachek argue, a la Becker (1985),that much of the remainingdiffer-
ential "probablyowes to the problemof controllingfor intensity of work
among individualswith greaterhome responsibilities"(p. 149).
Three English studies provide furtherevidence on the relationshipbe-
tween marital status and wages. Greenhalgh(1980), analyzing repeated

2. Although we focus on white women, we note that Hill finds positive associations between
both marriage and children and the earnings of black women.
3. Suter and Miller (1973) discuss unreported regression estimates from a cross-sectional
sample of women aged 30-44 drawn from the 1967 NLS Mature Women's file that lead
them to essentially the same conclusion as Hill: "It appears that once a woman's occupa-
tional status and work experience are known, learning that she is married and has children
does not significantly improve our ability to predict her income" (p. 192).
4. Goldin and Polachek also note (p. 149) that their results confirm those of an earlier study
by Polachek (1975) using 1960 Census data.

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
236 The Journal of Human Resources

cross-sections of British women from the GeneralHousehold Surveys of


1971 and 1975 finds mean hourly earnings differentialsfavoring single
women of 45 and 42 percent in 1971and 1975,respectively. But estimates
of separate wage equations by maritalstatus leave only 3 to 12 percent
of these differentials"unexplained"by differencesin the characteristics
of workers or jobs.
Siebertand Sloane (1981),using wage surveys of four Englishestablish-
ments, find roughly 10 to 25 percent unadjusted(i.e., mean) annualearn-
ings differentialsfavoring single women in three establishments,and an
11 percent differentialfavoringmarriedwomen in the fourth. Controlling
for workerattributes(typicallyexperience, tenure, and education)lowers
the differentialssubstantiallyin one establishment,but raises or leaves
unchangedthe others. Siebert and Sloane also report that the presence
of children under age 12 is unrelatedto wages for marriedwomen who
worked in the one establishmentthat collected informationon children.
Dolton and Makepeace (1987) estimate the association between mar-
riage, children, and wages using the (English) 1970Survey of Graduates.
The coefficient of a marriagevariable from an ordinary least squares
(OLS) log wage equation is -0.02 (with a standarderror of 0.02) for
women with no children, and -0.03 for those with children.5Correcting
for selectivity into employmentleads to a negligiblechange in the coeffi-
cient estimates.
In summary,these findingssuggest that marriagehas little or no associ-
ation with women's wages,6 while children appearto reduce wages pri-
marily "indirectly," by reducinglabor force participationand the accu-
mulation of human capital, rather than "directly," by lowering the
productivityof otherwise similarwomen. However, all of these studies
take childrenand maritalstatus, as well as experience and tenure, to be
exogenous determinantsof wage rates, and only one study (Dolton and
Makepeace)attempts to account for employmentselectivity.

III. Data and Empirical Findings

A. Data
The data analyzed are from the National LongitudinalSurvey of Young
Women. Most of our cross-sectional specificationsare estimated using
the 1982 wave of the survey, when the respondents were aged 28-38.

5. This last coefficient comes from an interactive specification; its standard error cannot
be computed from the information given in the paper.
6. For a review of the evidence for men see Korenman and Neumark (1991).

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
Korenman and Neumark 237

This period covers an age range that captures both post-schoolinglabor


market experience and marital status and fertility transitions that are
needed for longitudinalestimation. From the originalsample of 5,159 in
1968, attrition, nonmissing data requirements,and the restrictionof the
sample to white women reduce the sample size to 1,207 women working
for a wage as of the 1982 survey. To performlongitudinalanalyses, we
need repeated wage observations; much of our longitudinalanalysis fo-
cuses on the 911 women who worked for a wage in 1980 and 1982. We
will address the influenceof selection into the sample of women working
in 1982, as well as the smaller sample workingin both 1980and 1982.
In order to estimate a direct effect of marriage and children on
wages-apart from an indirect effect that may operate throughreduced
humancapital accumulation-good controls for labor marketexperience
and tenure are needed. Experience and tenure are constructedfromjob
history questions that were asked over the entire range (1968-82) of the
data set. Experience is measured as year-equivalentsof actual weeks
worked, while tenure simply counts the number of years for which a
respondentreports workingfor the same employer.

B. A First Look at Wage Levels and Changes


The upper panel of Table 1 reports mean log hourly wages (with wages
measuredin cents) for women with differentmaritalstatuses, and differ-
ent numbers of children, as well as the distributionof women across
these categories. Row (1) covers the sample of women working for a
wage in 1982. On average, never-marriedwomen earn wages 15 percent
higherthan divorced or separatedwomen who, in turn, earn slightlymore
than women who are marriedwith a spouse present.7More pronounced
differences appear between women with different numbersof children.
The greatest wage differential (about 27 percent) is found between
childless women and those with two or more children.
Although such differentialssuggest that marriageand childrenlower a
woman's wage, they may also reflect other differences (observable or
unobservable)among women. One approachto this heterogeneityprob-
lem is to examine how wages change as a woman changes maritalstatus
or has children.Rows (3) and (4) providea firstpass at this first-difference
analysis. Row (4) reports mean changes in log wages between 1980 and
1982 for women who changed maritalor fertility status. These changes
can be compared to the figures in Row (3), which are the mean changes
in log hourly earningsfor all women.

7. The category "divorced or separated"includes a few widows as well as a few women


who are marriedwith no spouse present.

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
Table 1
Mean Log Wages and Changes in Log Wages, for White Working Women Classified by M
of Children, and Changes in Marital Status and Number of Childrena

Status as of 1982
Married,
Spouse Divorced or Never- No
Present Separated Married Children

1982 sample
(1) Log wage 6.39 6.45 6.60 6.59
(1982) (.01) (.03) (.04) (.02)
N 787 262 158 371
1980-82 sample
(2) Log wage 6.46 6.47 6.59 6.60
(1982) (.02) (.03) (.04) (.02)

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
(3) Change in
log wage .20 .21 .22 .22
(1980-82) (.01) (.02) (.02) (.02)
N 576 194 141 316

Status Entered 1980-82


Married,
Spouse Divorced or One
Present Separated Child
1980-82 sample
(4) Change in
log wage .17 .21 .25
(1980-82) (.06) (.05) (.05)
Number of changers 46 53 33

a. Standard errors of means are reported in parentheses. Sample weights were not used in computing estimates.
for observations with wages and other variables used in wage regressions available for 1982. Wages are nominal
cents.

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
240 The Journal of Human Resources

The mean change in the log wage is lower for women who marry be-
tween 1980 and 1982 (0.17 in Row 4) than for women who remain single
in 1982 (0.22, Row 3). However, the mean change is only slightly lower
for women who have a second child (0.18, Row 4) compared to women
with one child as of 1982 (0.19, Row 3), and is actually relatively high for
women who have a first child between 1980 and 1982 (0.25, Row 4). Thus,
wage growth appears unaffected by changes in marital status or number
of children, raising the possibility that the cross-sectional differences in
wages by marital status or number of children may be mostly due to
heterogeneity, observable or not.
However, the first-difference approach may be flawed because recent
changers (e.g., women who had a first birth between 1980 and 1982) who
worked in 1982 may be a select group. In addition, because they are
aged 28-38 in 1982, these recent changers are also relatively "late" child
bearers, who tend to be relatively high earners (Bloom 1987, Blackburn
et al. 1990). Finally, as mentioned, the marital status and fertility transi-
tions may to some extent reflect. responses to wages. These potential
problems underscore the need for multivariate analyses that explicitly
account for heterogeneity, employment selectivity, and endogeneity of
marital status and number of children.

C. OLS Regression Estimates


Table 2 reports estimates from OLS cross-sectional log wage regressions.
Never-married women and childless women are the reference categories.
Column (1) reports the regression of log wages on the marital status
and number of children variables. As the estimates show, when dummy
variables for number of children are included in the equation, the negative
associations of marriage or divorce with wages (apparent in Table 1)
disappear, while children (especially two or more) are associated with
significantly lower wages. When years of education completed and
dummy variables for living in the South and in an SMSA are added in
Columns (2) and (3), the coefficients of the number of children dummy
variables are reduced by about one-half; only the coefficient of the two
or more children dummy variable remains statistically significant. In Col-
umn (4) experience and tenure are included, reducing the magnitude of
the coefficients of the number of children variables even further.8 These

8. Whenpotentialexperience (definedas age - schooling - six) is used in place of actual


experience and tenure, the maritalstatus coefficients are unchanged,while the children
coefficients are very close to those in Column (3) of Table 2. This is expected because
potentialexperience overstates actual labor marketexperiencethe most for women with
many laborforce interruptions,who are likely to be those with children.

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
Korenman and Neumark 241

Table 2
Wage Equation Estimates for White Working Women, 1982 Ordinary
Least Squares (dependent variable: natural logarithm of hourly
earnings)a

(1) (2) (3) (4)

Married, spouse present -.02 .01 .02 .05


(.05) (.04) (.04) (.04)
Divorced or separated -.00 .05 .05 .10
(.05) (.05) (.05) (.04)
One child -.13 -.05 -.05 -.04
(.04) (.04) (.04) (.04)
Two+ children -.30 -.18 -.18 -.07
(.03) (.03) (.03) (.03)
Education - .06 .06 .07
(.01) (.01) (.01)
South -- -.05 - .05
(.02) (.02)
Urban - .15 .15
(.03) (.02)
Experience - -- .02
(.004)
Tenure - - - .03
(.003)
F-testb .00 .00 .00 .03
R2 .10 .19 .23 .33

a. There are 1,207 observations.Standarderrorsare reportedin parentheses.Sample


weights were not used in computingestimates. Observationsare includedonly if the
wage reportedis for a job at which the respondentis currentlyworking.Never married
and no childrenare the referencecategories. Single-yearage dummyvariablesare in-
cluded in all specifications.
b. P-value for joint test of significanceof maritalstatus and fertilityvariables.

results are consistent with the findings of many of the studies reviewed
in Section II: after controlling for experience and tenure, marriage and
children have relatively little association with wages.9'10

9. We explored the sensitivity of the coefficientsreportedin Column(4) of Table 2 to the


inclusionof controlsfor, in turn,years marriedandyears divorcedor separated;the number
of preschool-agechildren;and age of the mother at first birth. The results do not differ
qualitativelyfrom those in Table 2: addingyears marriedand years divorcedor separated

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
242 The Journal of Human Resources

D. Bias in OLS Regression Estimates


As discussed in the introduction, there are many potential sources of
bias in the cross-sectional estimates presented in Table 2. The previous
subsection touched on biases that could be addressed by the inclusion
of better measures of observed characteristics. This section considers
sources of bias associated with unobservables: endogeneity; heterogene-
ity arising from selection into different categories of marital status and
number of children on the basis of unmeasured characteristics correlated
with wages; and selection into employment.

1. Endogeneity Bias and Instrumental Variables Estimates


Table 3 reports estimated coefficients and standard errors from equations
in which marital status and number of children, as well as experience and
tenure, are treated as potentially correlated with the wage equation error.
The table also reports test statistics for the exogeneity of these variables
with respect to the wage equation error (Hausman 1978). The instrumen-
tal variables, described in detail in the footnotes to Table 3, fall into two
categories: family background variables and measures of attitudes and
expectations. Exclusion restrictions to identify coefficients necessarily
involve untestable assumptions. In this section, the maintained assump-
tion is that family background measures are valid instruments. Research-
ers using sibling pairs to identify wage equation parameters have shown
that, once ability and schooling are taken in account, family background
does not have an independent effect on earnings or wages (Griliches
1979). These findings suggest using family background measures as instru-
ments. Conditional upon the maintained assumption that family back-
ground variables are valid instruments, the exclusion of measures of
attitudes and expectations from the wage equation can be tested as over-
identifying restrictions.

reducesthe coefficientsof the dummyvariablesfor maritalstatus, althoughonly the coeffi-


cient of the years divorced or separatedvariableis significant,and there is no negative
effect of marital status in any year of marriage;the effect of young childrenis slightly
positive, but insignificant;finally, the effect of childrenon wages varies with maternalage
at first birth, but the marital status and numberof children coefficients are essentially
unchanged.(These results are availableupon request.)
10. Perhapsthe most direct interpretationof Becker's hypothesisis that the joint effect of
marriage(or divorce) and childrenis to lower wages. The p-values for the F-statisticsfor
thejoint significanceof the maritalstatusandnumberof childrenvariablesin the regressions
in Table 2 indicate that these variables are jointly significant.However, in most of the
specificationsthe coefficientof the marriedvariableis positive. Indeed, in Column(4) the
point estimate of the summedeffect of marriageand childrenis close to zero: - 0.02 for
women with two or more children;and 0.01 for women with one child.

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
Korenman and Neumark 243

Instrumental variables (IV) estimates of the specifications from Table


2, Column (4)-using family background variables as instruments for ex-
perience, tenure, marital status and number of children-are reported in
Column (1) of Table 3. There is no statistical evidence that experience,
tenure, marital status, or number of children are correlated with the wage
equation error; in the bottom panel of the table, the p-values for the
exogeneity tests are all greater than 0.05.11 As the last entry in Column
(1) indicates, when the equation was reestimated including measures of
attitudes and expectations, the exclusion of these measures from the
wage equation could not be rejected. Hence, Column (2) adds attitudes
and expectations to the instrument list. In this specification there is also
no evidence that marital status or children are correlated with the wage
equation error; the p-values for the exogeneity tests are 0.79 and 0.27,
respectively. But the exogeneity of experience and tenure is rejected.
To examine the sensitivity of these conclusions to alternative specifi-
cations, in Columns (3) and (4) we specify a simpler wage equation that
includes a dummy variable for married, spouse present only (so that
never-married or divorced or separated is the reference category), and a
dummy variable for whether the woman has any children. Repeating the
analysis of Columns (1) and (2) with this simpler specification, our conclu-
sions are unchanged (using a 5 percent significance level). However, the
correlation between children and the wage education error is stronger;
the p-value for the exogeneity test in Column (4) is 0.07.
The instrumental variables methods do not lead to a rejection of the
exogeneity of marital status and children, and yield imprecise estimates
of their effects. Consequently, in Columns (5) and (6) we treat experience
and tenure (only) as potentially correlated with the wage equation error.
We continue to reject the exogeneity of experience and tenure, and their
coefficients are not significantly different from zero. For tenure this is
solely because the standard error increases, while for experience the
point estimate changes considerably (although a negative coefficient
seems implausible).12 More importantly, the coefficients of the marital
status and children variables are very similar to the OLS coefficients in
Column (3) of Table 2, in which experience and tenure were omitted.

11. This test involves: i) regressingthe potentiallycorrelatedvariableson the set of instru-


mentalvariablesand exogenous variables;ii) includingthe residualsfromthese regressions
in the wage equation estimatedby ordinaryleast squares;and iii) testing the joint signifi-
cance of the constructedresiduals.
12. These results contrastwith Heckman(1980)and Mincerand Polachek(1974),who find
no evidence of endogeneityof experiencein wage regressionsestimatedfor the NLS Mature
Women sample. They do not include tenure, and use a differentand more limited set of
instrumentalvariables(includingthe numberof young children).

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
Table 3
Wage Equation Estimates for White Working Women, 1982, Two Stage Least Squares (dep
natural logarithm of hourly earnings)a

Experience, Tenure, Marital Status,


and Fertility Endogenous
(1) (2) (3) (4)

Coefficients
Married, spouse present .40 -.02 -.07 -.14
(.51) (.30) (.31) (.1
Divorced or separated .75 .16 - -
(.59) (.30)
One child -.59 -.44 ~- -

(.69) (.41)
Two+ children -.38 -.40 ~- -

(.43) (.24)
- 27 -
Children .

(.36) (.2
Experience -.03 -.03 -.02 -.03
(.05) (.02) (.04) (.0
Tenure .03 .03 .01 .0
(.06) (.03) (.04) (.0

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
F-testb .59 .07 .49 .0
Family backgroundvariables used as IVsc Yes Yes Yes Ye
Expectational/attitudinalvariables used as IVsd No Yes No Ye
Specification tests (p-values)
Experience and tenure exogenouse .23 .00 .09 .0
Marital status exogenouse .38 .79 .80 .4
Fertility status exogenouse .59 .27 .54 .0
Expectational/attitudinal variables excludede .74 - .74 -

a. There are 1,207observations.Standarderrorsare reportedin parentheses.Sampleweights were not used in c


tions are includedonly if the wage reportedis for a job at which the respondentis currentlyworking.Single-yea
includedin all specifications.Never marriedand no childrenare the referencecategories. Otherindependentvar
in Table 2, Column(3).
b. P-value for joint test of significanceof maritaland fertilitystatus variables.
c. Variablesinclude:father's education;mother'seducation;parents'educationalgoal for respondentat age 14;
variableequal to one if the respondent'smotherworkedwhen respondentwas age 14; a dummyvariableequal to
with both a father and motherat age 14; and dummyvariablescorrespondingto each of these variables,equal to
missing(in which case the variableswere set equal to zero).
d. Variablesinclude:a dummyvariableset equal to one if respondentdisagreedor stronglydisagreedwith state
womanto work even if her husbanddisagrees,asked in 1971;a dummyvariableset equal to one if respondenta
this statement,in 1971;ideal age at marriagereportedby respondentat age 14 (set equal to zero, with a dummy
response was never to marry);expected numberof children,in 1970;educationalexpectations, in 1970;educatio
variablescorrespondingto each of these variables,equal to one when the variablewas missing (in which case th
zero).
e. P-value for joint significanceof coefficientsof the residualsfrom least squaresregressionof these variableson
variables,when these residualsare added to wage equationf. P-value for asymptoticF-test for exclusion of vari
timatedwith OLS.

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
246 The Journal of Human Resources

Of course, this finding stands in contrast to OLS estimates that include


experience and tenure, in which the negative association between chil-
dren and wages is attenuatedor eliminated.
As long as the error is uncorrelatedwith the instruments,the instru-
mental variables methods used in Table 3 are sensitive to any source of
correlationbetween the right-hand-sidevariablesand the wage equation
error. If the source of correlationis a fixed effect, then first-difference
methodsprovide an alternativeestimationstrategythat may be preferable
to IV methods for two reasons. First of all, first-differencemethods do
not rely on untestable exclusion restrictions. Second, if there are fixed
effects that, for example, influenceboth wages and fertility, IV methods
require not only an exclusion restriction, but an additionalassumption
that the excluded variable(s)are orthogonalto the fixed effect.

2. Unobserved Heterogeneity Bias and Fixed-Effects Estimates


We use first-differencesestimatedfor 1980and 1982to eliminatepotential
biases from fixed, unmeasured characteristics on the basis of which
women select into different maritalor fertility states. Although we ex-
clude experience and tenure from these specifications,this exclusion has
virtuallyno effect on estimates of the maritalstatus and childrencoeffi-
cients. This is because a short firstdifferenceimplicitlycontrolsfor expe-
rience and tenure by using a sample of women employedin two consecu-
tive survey years (e.g., 1980 and 1982). For women employed in two
consecutive years, having a child, for example, does not lead to much
reductionin experience and tenure, relative to childless women.13
For comparisonwith what follows, the firstrow of Table 4 reportsOLS
coefficient estimates for the subset of the samplefor which 1980and 1982
data are available.14The second row of the table reports estimates from
the 1980 and 1982 change equation.15None of the maritalstatus or chil-
dren coefficients is significantlydifferentfrom zero, nor are these vari-
ables jointly significant(the p-value from the F-test is 0.76). The coeffi-
cients of maritalstatus fall considerablyand become negative, while the
coefficients of the numberof childrenvariablesactuallybecome positive.

13. An implication of the fact that a short difference implicitly controls for experience and
tenure is that a short difference cannot be used to obtain an estimate of the direct plus the
indirect effect of marriage and motherhood on wages (i.e., effects that do not control for
experience and tenure).
14. We also carried out the analyses in Tables 2 and 3 using this smaller subsample, with
no changes in the conclusions.
15. The p-value for the asymptotic F-test of equality of coefficients across the 1980 and
1982 cross-sections-using the SUR estimator to account for cross-year correlations in the
residuals-was 0.93.

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
Table 4
Wage Equation Estimates for White Working Women, First-Difference Specifications (dep
and changes of natural logarithm of hourly earnings)a

Married,
Spouse Divorced or One Two + Ear
Present Separated Child Children Wa

1980-82 data
(1) 1982 .05 .05 -.01 -.15 -
(.04) (.05) (.04) (.04)
(2) 1982-80d change -.06 -.04 .05 .02 -
(.08) (.08) (.05) (.07)
Specification test
(3) 1982-80 change - .05 - .01 - .00 - .03 .04
(.07) (.08) (.05) (.07) (.03
Early difference
(4) 1973 .05 .06 .01 -.09 -
(.03) (.06) (.04) (.05)
(5) 1973-71d change .03 .07 -.02 -.02 -
(.03) (.06) (.06) (. 11)

a. Standarderrorsare reportedin parentheses.Sampleweights were not used in computingestimates. Observat


wage reportedis for a job at which the respondentis currentlyworking.Single-yearage dummyvariableswere i
for wage levels. Otherindependentvariablesare the same as those in Table 2, Column(3). Right-hand-sidevaria
and (4), and first differencesin all other rows. In the cross-sectionalspecifications,never marriedand no childre
b. 1978log wage.
c. P-value for joint test of significanceof maritalstatus and fertilityvariables.
d. Standardfixed-effectsestimator.

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
248 The Journal of Human Resources

The fixed-effects estimates suggest that cross-sectionalestimates are bi-


ased by unmeasuredheterogeneity. In other words, women with wage-
enhancingcharacteristics(net of observables) appearless likely to have
(two or more) children.
Heckman and Hotz (1989) have proposed an "overidentificationtest"
of the fixed-effect assumption. This test asks whether early wage levels
are associated with later wage growth, indicatingselection into different
maritalor fertility states based on wage growth, in contrast to the fixed-
effects assumptionof selection on wage levels. To performthis test, we
retain an earlier (1978) wage, and estimate
(w82 - w80) = (X82 - -
X80)b82 + w78d + (e82 e8o)
An estimated d significantlydifferentfrom zero indicates a violation of
the fixed-effectsassumption.The results reportedin Row (3) indicatethat
d is not significantlydifferent from zero, so we do not reject the fixed-
effects specification.16
Despite the results of the specificationtest, two problems may arise
from using 1980-82 first differences, because the effects of maritalstatus
and number of children are identifiedfrom women who changed states
between 1980 and 1982. First, because they are recent changers (i.e.,
they had a birth between 1980 and 1982), those who chose to work in
1982mightbe changerswith particularlyhigh wages. This is the standard
problem of selection into employment,but it may be particularlysevere
for recent changers (Solon 1988). We examine the problemof selectivity
of recent changers in the following subsection.
Second, the women in our sample were aged 26-36 in 1980, which
suggests that the changers are relatively late marriersor child bearers
who may have higher than average wages following first births or mar-
riage (perhapsbecause they are more able to afford child care), leading
to upward-biasedestimates in the first-differencespecifications. To ad-
dress the problem of "late changer" bias, we estimated an early first
difference (1971to 1973),when the women were 17 to 29 years old.17The
results, also reported in Table 4, differ little from the 1980-82 first-
difference estimates, indicatingthat there is no serious bias from using
late marriersand child bearers.
One approach to the problem of "recent changer" bias would be to
use a longer first difference, in which effects of childrenand marriageare
16. A similarspecificationtest of the fixed-effectsassumption(Heckmanand Robb 1986),
which uses informationfor two years priorto 1980and priorto marriageand childbearing,
also did not lead to rejectionof the fixed-effectsspecification.
17. The p-value for the test of equalityof coefficientsin the 1971and 1973cross sections
was 0.40. In addition, the 1971-73 first-differencespecificationwas also not rejected ac-
cordingto the overidentificationtest describedin the text.

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
Korenman and Neumark 249

identifiedfrom women who change states, even if they are not employed
in the years immediatelyfollowing the change. However, equality of the
wage equation coefficients for 1973 and 1982(the long difference we ex-
amined) was rejected (p - 0.02). Moreover, the fixed-effects specifica-
tion was rejectedaccordingto the overidentificationtest describedabove.
Thus, we could not use long differences to correct for selectivity among
recent changers. In the following section we use alternativemethods to
test and correct for employment selectivity.

3. EmploymentSelectivity Bias and Sample Selection Corrections


We first selectivity-correct the 1980-82 first-difference estimates,
applying Heckman's (1979) sample-selection correction techniques to
first differences, assuming jointly normally distributed errors. Before
turningto these estimates, OLS and selectivity-correctedestimates of the
1982cross-sectional wage equation are reportedin the first two columns
of Table 5.18 Since recent changers influence the cross-sectional esti-
mates, some selectivity bias in the cross-sectionalestimates is expected.
Indeed, the estimates of the numberof childrencoefficientsbecome more
negative when account is taken of selectivity, althoughthe changes are
small. The coefficient of the inverse Mills' ratio (lambda)is marginally
significant,with a t-statistic of 1.86.
In Columns (3) and (4) we apply a similar technique to the 1980-82
first difference, using a bivariate selectivity criterionfor employment in
1980 as well as 1982.19The coefficient estimates are unaffected by the
selectivity correction, and the estimatedcoefficients of the lambdaterms
(one for each year) do not indicate the presence of sample selectivity.
As an alternative test for selectivity bias from recent changers-one
that does not requirethe specificationof employmentequations-we esti-
mate a wage regression for all women who worked in 1980 (whether or

18. Variablesincluded in the employmentprobit, but excluded from the wage equation,
include:husband'sincome; income from alimonyor child support;and weeks the husband
spent unemployedin the year precedingthe 1982 survey. The sample is reduced slightly
because these additionalvariables used in the employmentprobit were occasionally un-
available.
While these variables should affect the reservationwage, and not the offer wage, they
are correlatedwith maritalstatus, and may thereforeprovidelittle independentinformation.
However, because underthe normalityassumptionthe selectionmodel is identifiedwithout
these exclusion restrictions(Olsen 1980), the restrictionswere tested by includingthese
variablesin the wage equationand testing their significance;the p-value for the F-test of
the joint significanceof these three variablesin the wage equationwas 0.59.
19. Details are given in the footnotes to Table 5. The exclusion restrictionof husband's
income, income from alimony or child support, and weeks unemployed(of the husband)
was not rejected (p = 0.78).

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
250 The Journal of Human Resources

Table 5
Wage Equation Estimates for White Working Women, Sample
Selectivity-Corrected (SSC) Estimates (dependent variable levels and
changes of natural logarithm of hourly earnings)a

1982 1982 1982-80 1982-80


OLS SSCb1 OLSC SSCd
(1) (2) (3) (4)

Status in 1982
Married, spouse present .02 .01 -.05 -.07
(.04) (.04) (.08) (.08)
Divorced or separated .05 .07 -.04 -.05
(.05) (.05) (.08) (.08)
One child -.05 -.07 .04 .05
(.04) (.04) (.05) (.06)
Two or more children -.18 -.22 .04 .05
(.03) (.04) (.07) (.09)
Lambda - .13 -
(.07)
Lambda-80 - - - -.04
(.05)
Lambda-82 - .06
(.07)
F-teste .00 .00 .83 .75
N 1,181 1,181 888 888

a. Standard errors are reported in parentheses. Sample weights were not used in comput-
ing estimates. Observations are included only if the wage reported is for a job at which
the respondent is currently working. Single-year age dummy variables were included in
all specifications for wage levels. Other independent variables are the same as those in
Table 2, Column (3). In the cross-sectional specifications, never married and no children
are the reference categories.
b. Variables in the employment probit include 1982 values of all variables included in
wage equation specification in Table 2, Column (3), as well as dummy variables indicat-
ing whether a respondent changed marital or fertility states between 1980 and 1982
(a different dummy variable is used for each of the four categories). Finally, measures
of husband's income and weeks husband spent unemployed in 1982 (both set to zero for
unmarried women), and the sum of income from alimony and child support (set to zero
for never-married women), are included.
c. Standard fixed-effects estimator.
d. A bivariate probit model is used for the selectivity correction. Variables included in
probits for 1982 and 1980 are values for corresponding year of all variables included in
wage equation specification in Column (3) of Table 2, values of husband's income and
weeks husband spent unemployed (both set to zero for unmarried women), the sum of in-
come from alimony and child support (set to zero for never-married women), and-for
1982-dummy variables indicating whether a respondent changed marital or fertility
states between 1980 and 1982 (a different dummy variable is used for each of the four
categories).
e. P-value for joint test of significance of marital status and fertility variables.

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
Korenman and Neumark 251

not they worked in 1982). In this wage regression for 1980, we include
dummy variables indicating whether a woman changed into one of the
maritalor fertility states between 1980 and 1982, as well as interactions
between these variables and a variable indicating whether the woman
was employed in 1982.The coefficients of the interactionterms measure,
for example, the wage in 1980of a womanwho had her firstchild between
1980and 1982and workedin 1982, relativeto the wage of a similarwoman
who had her first child between 1980and 1982, but was not employed in
1982. If high earnersamongrecent changersselect into employment,then
a woman who had a child between 1980 and 1982 and who worked in
1982, should earn relatively higherwages in 1980than a woman who had
a child between 1980 and 1982but did not work in 1982.
Table 6 reports results for these specificationsfor both 1971and 1980,
to examine the importanceof selectivity bias in the 1971-73 and 1980-82
first differences. We find no evidence of selectivity bias from recent
changers. None of the coefficients of the recent changer interactions is
statistically significant. These conclusions do not differ from the tests
based on the more standardselectivity-correctionmethods.

IV. Conclusion
We have explored the consequences of a numberof poten-
tial problems with drawing causal inferences from cross-sectional rela-
tionships between marriage, motherhood, and wages. These problems
include: endogeneity of marriageand motherhood, and experience and
tenure; heterogeneity;and selectivity into employment.
We have three main findings to report. First, introducingexperience
and tenure into wage equations estimated by OLS attenuates but does
not eliminatethe large negative relationshipbetween childrenand wages.
Instrumentalvariables techniques and the accompanyingtests suggest
that marital status and number of children are exogenous variables in
wage equations. However, IV results also suggest that experience and
tenure are not exogenous. This findingis importantbecause the size and
statistical significance of the wage effects of children in cross-sections
are sensitive to the exclusion of experience and tenure controls, and to
their estimated coefficients. Instrumentingfor experience and tenure
yields estimated effects of maritalstatus and numberof childrenthat are
similar to OLS estimates when experience and tenure are excluded; in
particular,the negative effects of having two or more childrenappearto
be large.
Second, first-differencespecificationssuggest that fixed unobservables
bias cross-sectional estimates of the effects of childrenon wages. Short

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
Table 6
Wage Equation Estimates for White Working Women, Survey Year Prior to Change in Ma
of Children (dependent variable: natural logarithm of hourly earnings)a

1980 Wage Regression

Became Became
Married, Divorced or
Spouse Present Separated F

(Change in marital or fertility status) x .02 .24


(worked in 1982)b (.15) (.27)
Number worked 1980 and 1982C 50 57
Number worked 1980 onlyd 7 2

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
1971 Wage Regression

(Change in marital or fertility status) x .02 -.17


(worked in 1973)b (.05) (.11)
Number worked 1971 and 1973c 111 20
Number worked 1971only d 65 9

a. There are 833 observationsfor the 1971regression,and 1,091 observationsfor the 1980regression. Standarde
theses. Sampleweights were not used in computingestimates. For each year, observationsare included if the wa
which the respondentis currentlyworking.Single-yearage dummyvariableswere includedin all specificationsf
dent variableslisted in Table 2, Column(3) are included.In addition,dummy-variableindicatorsof changes in m
childrenare included.
b. Estimatedcoefficientsof variablesdefinedas the 1982(1973)maritalstatus or numberof childrenvariable, tim
to one if the woman changedinto the categorybetween 1980and 1982(1971 and 1973),times a dummy variable
mainedin the workforcein 1982(1973).
c. Numberof women who changedinto the indicatedmaritalstatus or numberof childrencategory between 198
and workedfor a wage in 1982(1973).
d. Numberof women who changedinto the indicatedmaritalstatus or numberof childrencategory between 198
and did not work for a wage in 1982(1973).

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
254 The Journal of Human Resources

first-differences (estimated over a two-year period) indicate no negative


effects of motherhood on wages.
Finally, standard sample-selection corrections, as well as selectivity
tests that do not depend upon specifying an employment equation, pro-
vide no evidence of selectivity bias from using a sample of recent chang-
ers (women who are employed despite recent changes of marital or fertil-
ity states).
Like many previous cross-sectional estimates, the first-difference esti-
mates do not support the contention that marriage or motherhood lower
women's wages. However, the IV results call into question specifications
that include controls for experience and tenure, either explicitly (as in
cross-sections) or implicitly (as in first differences). Estimates from such
specifications are likely to understate the direct effects of children on
wages, because the lower experience and tenure associated with marriage
and motherhood may arise as an endogenous response to lower wages.
In other words, the IV results suggest that children lower wages "di-
rectly," and women respond to these lower wages by curtailing their
labor supply and hence the accumulation of labor market experience and
tenure. This conclusion contrasts with the more standard conclusion-
suggested by our fixed-effects estimates and by estimates from earlier
studies-that children lower wages primarily "indirectly" by lowering
attachment to the labor force, or experience and tenure. If the standard
conclusion is correct, then by remaining employed women would experi-
ence no adverse wage effects from having children. On the other hand,
the IV estimates suggest that high labor force attachment alone will not
prevent child bearing and child rearing from having an effect on women's
wages. The sensitivity of the principal conclusions to the treatment of
experience and tenure, therefore, highlights the need for continued re-
search on the more general question of whether experience and tenure
should be treated as exogenous variables in wage equations for women.

References
Becker, G. S. 1981.A Treatiseon the Family. Cambridge:HarvardUniversity
Press.
. 1985. "HumanCapital,Effort, and the Sexual Division of Labor."
Journal of Labor Economics 3(1, Pt. 2):S33-S58.
Blackburn,M. L., D. E. Bloom, and D. Neumark. 1990. "FertilityTiming,
Wages, and Human Capital." NBER WorkingPaper.
Bloom, D. E. 1987. "FertilityTiming, Labor Supply Disruptions,and the Wage
Profiles of American Women." 1986 Proceedings of the Social Statistics
Section of the American Statistical Association: 49-63.
Butz, W. P., and M. P. Ward, 1979. "The Emergenceof CountercyclicalU.S.
Fertility." American Economic Review 69(3):318-28.

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions
Korenman and Neumark 255

Dolton, P., and G. Makepeace. 1987. "MaritalStatus, ChildRearingand


EarningsDifferentialsin the GraduateLabourMarket." The Economic
Journal 97(388):897-922.
Easterlin, R. A. 1980. Birth and Fortune: The Impact of Numbers on Personal
Welfare.New York: Basic Books, Inc.
Fuchs, V. 1989. "Women's Quest for Economic Equality." Journal of
Economic Perspectives 3(1):25-41.
Goldin, C., and S. Polachek. 1987. "ResidualDifferencesby Sex: Perspectives
on the Gender Gap in Earnings. "American Economic Review 77(2):143-51.
Greenhalgh,C. 1980. "Male-FemaleWage Differentialsin Great Britain:Is
Marriage an Equal Opportunity?" The Economic Journal 90(360):751-75.
Griliches, Z. 1979. "Sibling Models and Data in Economics: Beginningsof a
Survey." Journal of Political Economy 87(5, Pt. II):S74-103.
Hausman,J. A. 1978. "SpecificationTests in Econometrics." Econometrica
46(6):1251-71.
Heckman, J. J. 1979. "Sample Selection Bias as a SpecificationError."
Econometrica 47(1):153-61.
. 1980. "Sample Selection Bias as a SpecificationErrorwith an
Applicationto the Estimationof Labor Supply Functions." In Female Labor
Supply, ed. J. P. Smith, 206-48. Princeton:PrincetonUniversity Press.
Heckman, J. J., and V. J. Hotz. 1989. "ChoosingAmong Alternative
NonexperimentalMethods for Estimatingthe Impactof Social Programs:
The Case of Manpower Training." Journal of the American Statistical
Association 84(408):862-74.
Heckman, J. J., and R. Robb. 1986. "AlternativeIdentifyingAssumptionsin
EconometricModels of Selection Bias." Advances in Econometrics
5:243-87.
Hill, M. 1979. "The Wage Effects of MaritalStatus and Children."The
Journal of Human Resources 14(4):579-94.
Korenman,S., and D. Neumark. 1991. "Does MarriageReally Make Men
More Productive?" Journal of Human Resources 26(2):282-307.
Mincer, J., and S. Polachek. 1974. "Family Investmentsin HumanCapital:
Earningsof Women." In Economics of the Family, ed. T. W. Schultz,
397-429. Chicago: The University of ChicagoPress.
Moore, W., and R. Wilson. 1982. "The Influenceof Childrenon the Wage
Rates of MarriedWomen." Eastern EconomicJournal 3(3):197-210.
Olsen, R. 1980. "A Least Squares Correctionfor Selectivity Bias."
Econometrica 148(7):815-20.
Polachek, S. 1975. "Differences in Post-School Investmentas a Determinantof
Market Wage Differentials." International Economic Review 16(2):451-70.
Siebert, W., and P. Sloane. 1981. "The Measurementof Sex and MaritalStatus
Discriminationat the Workplace."Economica 48(190):125-41.
Solon, G. 1988. "Self-Selection Bias in LongitudinalEstimationof Wage
Gaps." Economics Letters 28(3):285-90.
Suter, L. E., and H. P. Miller. 1973. "Income Differences Between Men and
Career Women." American Journal of Sociology 78(4):962-75.

This content downloaded from 178.250.250.21 on Fri, 13 Nov 2015 07:17:42 UTC
All use subject to JSTOR Terms and Conditions